acceptodds
Under review as a conference paper at ICLR 2027

HalluMirror: Discovering and Mitigating Multimodal Hallucinations via Co-Evolutionary Self-Play

Abstract

Mitigating hallucinations in Multimodal Large Language Models (MLLMs) increasingly relies on constructed training supervision. However, existing methods typically optimize against predefined hallucination patterns, curated preferences, or model-generated samples whose challenges are not explicitly adapted to the model's evolving failure modes. We identify this mismatch as a key limitation: as the model improves, effective supervision should continually shift toward its current unresolved hallucination weaknesses. To this end, we propose HalluMirror, a co-evolutionary adversarial self-play framework that continually transforms model failures into new training challenges. HalluMirror couples an Inducer, optimized to generate fine-grained existence, attribute, and relation hallucinations that expose the current Challenger's weaknesses, with a Challenger that learns to detect and repair these evolving hard cases. Through alternating optimization, weakness discovery, adversarial exposure, and targeted repair form a closed self-improvement loop. To reliably scale this process, we further introduce a training-free grounding-based hallucination verifier that filters self-play-generated hallucinations without hallucination-specific human annotations. Extensive experiments across multiple MLLMs and hallucination benchmarks show that HalluMirror consistently improves robustness against diverse hallucination types. On GLM-4.1V-9B, HalluMirror improves performance across all eight hallucination benchmarks, with co-evolution identifying and addressing the model’s vulnerability to relation hallucinations and yielding a 14.5% relative gain on FINER’s multi-relation subset. All data and code will be made publicly available.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
MLLM hallucination mitigation, self-evolution

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ia2zOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HalluMirror, a novel self-play framework for mitigating hallucinations in MLLLMs. An Inducer generates visual inconsistencies tailored to a Challenger's evolving weaknesses, while the Challenger learns to detect and repair them. A training-free verifier filters generated examples. Several experiments showed improved hallucination robustness. Especially, the proposed method improved performance across all eight hallucination benchmarks on GLM-4.1V-9B.

Critical strengths
  1. The paper addresses an important problem in mitigating hallucinations in MLLMs.
  2. The methodology sounds good and the methodological contribution is good for ICLR.
  3. A gap is addressed in adapting hallucination mitigation to the evolving weaknesses of MLLMs.
Critical weaknesses
  1. The performance gain is insufficient. The proposed method underperformed existing methods on the majority of benchmarks (win:loss=5:6) in Table 1.
  2. The proposed method should be compared with APASI[1] in Table 1.
  3. The hallucination verifier evaluation lacks important baselines[2][3].
  4. The motivation for only using existence/attribute/relation is unclear. I guess that the limited performance on FREAK and MMHal-Bench may stem from this restricted scope. Previous works [2][3][4] have considered additional hallucination types.
Questions for authors

See the critical weakness.

Reviewer_cNxBOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

HalluMirror uses adversarial self-play to reduce multimodal hallucinations. An Inducer learns to introduce plausible but visually incorrect edits into image descriptions, while a Challenger learns to detect and repair them. The two models are updated alternately, with a grounding-based verifier and difficulty-based filtering used to select training examples. Experiments on three MLLMs show broad improvements across hallucination benchmarks, with generally small changes in general capabilities.

Critical strengths

The method provides a concrete way to adapt hallucination training to the model’s current weaknesses. The analysis of attack success, detection, and repair across iterations helps explain how the two roles interact. The improvement on FINER’s relation subset provides useful evidence that the benefits extend beyond the self-play examples.

The evaluation is reasonably broad, covering three model families, eight hallucination benchmarks, and general capability checks. The fixed-data GRPO comparison also matches the cumulative data size and hallucination-count distributions, making it more informative than a comparison against the base model alone.

Critical weaknesses

The ablation does not fully isolate the benefit of learning an adaptive Inducer. Comparing against a fixed dataset changes both the generator and whether the current Challenger refreshes and selects examples. A frozen Inducer that generates fresh examples each round, with the same difficulty filtering, would help establish whether adversarially updating the Inducer is necessary.

The verifier is evaluated on standard benchmarks, but its reliability on the evolving adversarial examples remains unclear. These examples may become increasingly difficult for the verifier as well as the Challenger. Since accepted edits determine the training labels, verifier mistakes could introduce incorrect supervision. A manual audit of accepted examples across iterations would make the self-play results more convincing.

The related work lacks discussion of UniGame and VILLA, which explore multimodal adversarial self-play and embedding-space adversarial perturbations, respectively. Discussing how HalluMirror differs in its perturbation mechanism and training objectives would help clarify its contribution.

Questions for authors

Has the author manually checked the hallucination labels in the filtered training data? Does their accuracy change across self-play iterations?

Reviewer_sNfROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

The paper introduces HalluMirror, a method that mitigates hallucinations in MLLMs in an evolving manner. They propose two models, an Inducer and a Challenger. The Inducer is fine-tuned during the process to generate hallucinations that can fool the Challenger, while the Challenger is fine-tuned to detect and correct the hallucinations generated by the Inducer. They also introduce a grounding-based hallucination verifier to filter invalid hallucinations from the generated training data. After three iterations of this process, they evaluate their method on various hallucination benchmarks and different MLLMs, showing improvements across most benchmarks. They also conduct an ablation study to investigate the effect of their evolving training process compared to fixed-data training and separately evaluate the effectiveness of their hallucination verifier. Their results suggest that the proposed method can adapt to the weaknesses of different models, with particularly noticeable improvements in relation hallucinations.

Critical strengths
  1. The paper has an interesting motivation to reduce hallucinations in an evolving manner.

  2. The authors conduct extensive experiments to show the effectiveness of their method, including evaluations on multiple benchmarks and models, as well as ablation studies.

  3. The method itself is interesting, as it uses two versions of a model that interact with each other and improve over time.

  4. The evaluation covers different types of hallucination. Also, they have evaluated the general capability of the models after fine-tuning to make sure the method does not hurt the overall performance of the MLLMs.

Critical weaknesses
  1. The paper is quite difficult to follow, and I believe the organization of the method could be improved. For example, in Section 2.1, the authors describe how the Inducer exposes the Challenger's weaknesses and how the generated rollouts are evaluated and selected, without clearly explaining how these processes work. The reward functions are introduced after the data selection process, making the overall training procedure harder to understand. Also, the authors repeatedly use abstract terms such as "evolving weaknesses" and "adversarial exposure" without clearly explaining them.

  2. The novelty of the proposed method is not very clear to me. The idea of having two evolving models has already been explored in previous self-play methods. I understand that this paper extends this idea to multimodal hallucination detection and correction, but it is not clear what is fundamentally new beyond this adaptation.

  3. The main motivation of the paper is that the Inducer continuously discovers new weaknesses of the Challenger. However, I do not think the experiments clearly support this argument. It is not clear whether updating the Inducer is necessary or whether the same Inducer could still generate useful examples for the Challenger in later iterations. I believe an ablation where the Inducer is kept frozen but new samples are generated using the same strategy at each iteration would better support this claim. Also, providing some qualitative examples showing how new weaknesses are discovered and resolved across iterations would be beneficial.

Questions for authors

Why is it necessary to update the Inducer at each iteration? Could the same Inducer still generate useful examples for the updated Challenger? It would be interesting if the authors could design an experiment or provide some qualitative examples to show that new weaknesses are actually being discovered and resolved across iterations.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all reviewers for their careful evaluation, constructive feedback, and thoughtful consideration. HalluMirror addresses two connected challenges in multimodal hallucination mitigation: adapting training signals to a model’s evolving hallucination weaknesses and validating automatically generated supervision against visual evidence. Below, we summarize our core contributions and the additional evidence provided during the rebuttal.

Core Contributions

  1. Adaptive Discovery and Repair of Hallucination Weaknesses. HalluMirror couples an Inducer and a Challenger through alternating adversarial self-play. The Inducer learns to generate fine-grained hallucinations targeting the current Challenger’s failures, while the Challenger learns to detect and repair them. This interaction continually converts unresolved hallucination weaknesses into targeted training signals.

  2. Grounding-Based Verification for Self-Play. We introduce a training-free verifier that checks existence, attribute, and relation hallucinations against visual evidence. Together with injection records and faithful reference descriptions, this supports the automatic construction of detection-and-repair supervision without hallucination-specific human annotations.

  3. Broad Empirical Validation. Experiments cover three MLLM families and eight hallucination benchmarks. On GLM-4.1V-9B, HalluMirror improves over the base model across all eight benchmarks after three iterations, using at most 1.05K curated Challenger training examples per iteration. Evolution analyses further connect model-specific weaknesses with targeted improvements, including a 14.5% relative gain on FINER’s multi-relation subset.

Additional Evidence During the Rebuttal

  • Controlled validation of Inducer adaptation: We added a frozen-Inducer baseline with fresh samples generated each iteration, matching data volume, difficulty filtering, and other training settings. HalluMirror achieves a higher final average score (63.98 vs. 61.59), supporting the benefit of adapting the Inducer to the current Challenger.
  • Supervision reliability analysis: We manually audited verifier-accepted samples across all three iterations, quantifying supervision validity and analyzing verification reliability during self-play.
  • Additional comparisons and qualitative evidence: We added an APASI-IACL comparison using SI-23K, further verifier comparisons with HalLocalizer and ZINA under the settings detailed in our responses, and qualitative examples illustrating weakness discovery and repair across iterations.

We again sincerely thank all reviewers for their time, valuable suggestions, and thoughtful consideration.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.